U.S. patent application number 14/542202 was filed with the patent office on 2016-05-19 for managing skew in data signals with multiple modes.
The applicant listed for this patent is Cavium, Inc.. Invention is credited to David Da-Wei Lin, Edward Wade Thoenes.
Application Number | 20160141018 14/542202 |
Document ID | / |
Family ID | 55962276 |
Filed Date | 2016-05-19 |
United States Patent
Application |
20160141018 |
Kind Code |
A1 |
Lin; David Da-Wei ; et
al. |
May 19, 2016 |
MANAGING SKEW IN DATA SIGNALS WITH MULTIPLE MODES
Abstract
A method for controlling a memory includes causing a data
de-skewer to operate in a writing mode, at the data de-skewer,
receiving a first signal, and skewing the first data signal by a
first compensation skew, causing the data de-skewer to operate in a
reading mode, at the data de-skewer, receiving a second signal, and
skewing the second signal by a second compensation skew, wherein
the first signal is representative of a bit from a byte that is to
be written to the memory, and wherein the second signal is
representative of a bit from a byte that has been read from the
memory.
Inventors: |
Lin; David Da-Wei;
(Westborough, MA) ; Thoenes; Edward Wade;
(Cambridge, MA) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Cavium, Inc. |
San Jose |
CA |
US |
|
|
Family ID: |
55962276 |
Appl. No.: |
14/542202 |
Filed: |
November 14, 2014 |
Current U.S.
Class: |
365/193 ;
365/191 |
Current CPC
Class: |
G11C 7/1093 20130101;
G06F 13/1689 20130101; G11C 2207/2254 20130101; G11C 11/4094
20130101; G11C 11/4076 20130101 |
International
Class: |
G11C 11/4094 20060101
G11C011/4094; G11C 11/4076 20060101 G11C011/4076 |
Claims
1. An apparatus for controlling a memory, said apparatus
comprising: a memory controller, and an interface to data lines
connecting said memory controller to said memory, wherein each of
said data lines carries a signal that corresponds to a bit of a
byte that is to be written to said memory, wherein said interface
comprises, for each of said data lines, a transmitter to drive bits
to be written into said memory onto a data line coupled to the
transmitter, wherein said interface comprises, for each of said
data lines, a receiver to detect bits that have been read from said
memory from a data line coupled to the receiver, wherein said
memory controller comprises, for each of said data lines, a data
de-skewer used in a writing mode and a reading mode, wherein for
each of said data lines, said data de-skewer is configured to
receive a first data signal, wherein each of said data lines is
associated with an inherent skew, wherein said data de-skewer
applies a compensation skew to said first data signal to generate a
second data signal, and wherein, in the writing mode said first
second data signal is a signal that represents a bit that is to be
written into said memory, and in the reading mode said second data
signal is a signal that represents a bit that has been read from
said memory.
2. The apparatus of claim 1, wherein, for each of said data lines,
said compensation skew applied to said first data signal to
generate a second data signal is obtained by training said data
de-skewer in the reading mode.
3. The apparatus of claim 2, wherein, for each of said data lines,
said data de-skewer uses said compensation skew, which was obtained
by training said data de-skewer in the reading mode, for skewing a
data signal that represents a bit that is to be written into
memory.
4. The apparatus of claim 2, wherein said data de-skewer is trained
in the reading mode using data signals that represent bits provided
by said memory during normal operation of said memory.
5. The apparatus of claim 2, wherein said data de-skewer is trained
based on locating timing of a leading-transition interval and
trailing-transition interval of data signals representing bits
propagating over each data line.
6. The apparatus of claim 5, wherein locating timing of a
leading-transition interval and a trailing-transition interval of a
first data signal includes oversampling said first data signal
based on a strobe signal corresponding to said first data
signal.
7. The apparatus of claim 6, wherein oversampling said first data
signal based on the corresponding clock signal includes sampling
said first data signal based on the corresponding strobe signal and
multiple time-shifted versions of the corresponding strobe
signal.
8. The apparatus of claim 1, wherein, for each of said data lines,
said data de-skewer uses a first input value for determining the
compensation skew in the reading mode, and uses a second input
value, different from the first input value, for determining the
compensation skew in the writing mode.
9. The apparatus of claim 8, wherein, for each of said data lines,
said second input value is obtained by training said data de-skewer
by writing test data into said memory in the writing mode and
reading said test data from said memory in the reading mode using
said compensation skew determined in the reading mode.
10. The apparatus of claim 8, wherein, for each of said data lines,
said first input value is obtained by training said data de-skewer
based on locating timing of a leading-transition interval and
trailing-transition interval of a data signal representing bits
propagating over said data line.
11. A method for controlling a memory, said method comprising:
causing a data de-skewer to operate in a writing mode, at said data
de-skewer, receiving a first signal, and skewing said first data
signal by a first compensation skew, causing said data de-skewer to
operate in a reading mode, at said data de-skewer, receiving a
second signal, and skewing said second signal by a second
compensation skew, wherein said first signal is representative of a
bit from a byte that is to be written to said memory, and wherein
said second signal is representative of a bit from a byte that has
been read from said memory.
12. The method of claim 11, further comprising training said
de-skewer while said memory is operating in said reading mode.
13. The method of claim 12, wherein training said de-skewer
comprises generating said first compensation skew.
14. The method of claim 12, wherein training said data de-skewer in
reading mode comprises using data signals that represent bits
provided by said memory during normal operation of said memory.
15. The method of claim 12, wherein training said data de-skewer in
reading mode comprises locating a leading-transition interval and a
trailing-transition interval of a data signal that represents a bit
that is propagating over a data line.
16. The method of claim 15, wherein locating said
leading-transition interval and said trailing-transition interval
of said data signal comprises oversampling said data signal based
on a clock signal, wherein said clock signal is a clock signal that
corresponds to said data signal.
17. The method of claim 16, wherein oversampling said data signal
based on said corresponding clock signal comprises sampling said
first data signal based on a corresponding clock signal and on
multiple time-shifted versions of said corresponding clock
signal.
18. The method of claim 11, wherein, for each of a plurality of
data lines, causing said data de-skewer to use a first input value
for determining a compensation skew in said reading mode, and
causing said data de-skewer to use a second input value, different
from said first input value, for determining a compensation skew in
said writing mode.
19. The method of claim 18, wherein, for each of said data lines,
obtaining said second input value by training said data de-skewer,
wherein training said data de-skewer comprises writing test data
into said memory in said writing mode, and reading said test data
from said memory in said reading mode using said compensation skew
determined in said reading mode.
20. The method of claim 19, wherein, for each of said data lines,
obtaining said first input value by training said data de-skewer
based on locating timing of a leading-transition interval and
trailing-transition interval of a data signal representing bits
propagating over said data line.
21. The method of claim 11, further comprising selecting said first
compensation skew and said second compensation skew to be equal to
each other.
22. The method of claim 11, further comprising selecting said first
compensation skew and said second compensation skew to be different
from each other.
23. The apparatus of claim 1, wherein said data de-skewer comprises
a delay line that is used for both writing mode and reading mode.
Description
BACKGROUND
[0001] This application relates to memory controllers, and in
particular, to managing skew in data signals, using multiple
modes.
[0002] In a typical memory system, a memory and a memory controller
connect to a common bus that has several parallel wires. Each wire,
which carries one bit, will be referred to herein as a "data
line."
[0003] The signal present on each data line represents data. It
will therefore be referred to herein as the "data signal." In a
typical memory with eight-bit bytes, there would be eight of these
data signals for every byte.
[0004] In a typical memory system, the wires that carry a byte
include data lines, with one data line for each bit. Thus, for an
eight-bit byte, there would be eight data lines. However, there is
also an extra pair of wires. This extra wire pair carries a
differential signal that functions as a reference clock for
sampling the data lines. The differential signal carried by these
extra wires is often called a "strobe." The wires that carry the
strobe are collectively called a "strobe line."
[0005] For example, such a memory system may include a memory
controller for a memory module that includes a number of
synchronous dynamic random access (SDRAM) memory chips. In such a
system the data line may be called a "DQ" signal, and the strobe
may be called a "DQS" signal.
[0006] In the process of writing to the memory, a memory controller
receives several n-bit units worth of data, together with an
instruction to write those units into the memory. Each unit has n
data bits and an associated strobe line. For each n-bit unit, the
controller places each of the n bits on each of the n individual
data lines corresponding to those n bits. As part of this task, the
memory controller places a strobe on the strobe line for each n-bit
unit. For each n-bit unit, the memory makes use of the strobe to
sample the data lines for that unit, and stores the corresponding
data into the memory array. In some embodiments, n=8. However, in
other embodiments n=4.
[0007] The data lines may be bidirectional. So, in the process of
reading from the memory, the controller issues one or more commands
to the memory. In response, the memory places data signals on the
data lines and a corresponding strobe on the strobe line. The
controller then retrieves the data from the bidirectional data
lines.
[0008] In the ideal case, the timing of the strobe and the data
signals should be maintained with a particular relative phase
offset while propagating between the controller and the memory.
However, in practice, different data lines have different skews in
the data signals. These different skews, which result in different
relative phase offsets, arise as a result of variations in the
path-lengths of the routes followed by the data signals as they
traverse the package that contains a memory component, a module
containing multiple packaged memory components, a printed circuit
board that connects the controller to the memory module, and the
package of the memory controller. For example, temperature may also
cause such variations. As a result of higher data rates, the
effects of these variations are more significant.
SUMMARY
[0009] In one aspect, the invention features an apparatus for
controlling a memory. Such an apparatus includes a memory
controller, and an interface to data lines connecting the memory
controller to the memory. Each of the data lines carries a signal
that corresponds to a bit of a byte that is to be written to the
memory. The interface includes, for each of the data lines,
circuitry for transmission of a bit to be written into the memory
via the data line, and for each of the data lines, a data
de-skewer. For each of the data lines, the data de-skewer is
configured to receive a first data signal that represents a bit
that is to be written into the memory. Each of the data lines is
associated with an inherent skew. The data de-skewer applies a
compensation skew to the first data signal to generate a second
data signal. An extent of this compensation skew is selected to
increase a likelihood that the second data signal will be sampled
by the memory during a data-valid window thereof. The same data
de-skewer is further configured to receive a first data bit that
has been read from the memory and to skew the first data bit.
[0010] In some embodiments, a sum of the extent of the compensation
skew and the inherent skew of the first data signal equals a total
skew. The extent is selected to minimize a difference between the
total skew and a skew that is common to all bits in the byte.
[0011] Other embodiments include a strobe de-skewer configured to
receive a first strobe and to output a second strobe. The second
strobe is to be placed on a strobe line. The strobe de-skewer is
configured to skew the first strobe to an extent selected to
maximize a likelihood that the second strobe will cause sampling of
all data bits on the data lines during a data-valid window that is
common to all of the data bits on the data lines. Among these
embodiments are those in which the strobe de-skewer is configured
to skew the first strobe by a value that is midway between a
highest and lowest skew value of the strobe de-skewer. In some
cases, the highest and lowest skew value of the strobe de-skewer
and a highest and lowest skew value of the data de-skewer are the
same.
[0012] Further embodiments include those that include a multiplexer
having an output that is connected to an input of the data
de-skewer. This multiplexer has a first input and a second input.
The first input is connected to a portion of the interface
configured to receive, from the data line, a signal indicative of a
bit that has been read from the memory. The second input is
configured to receive the first data signal. The de-skewer has an
output that is selectively connected to the data line so that the
second data signal can be placed on the data line. The choice of
what input to select depends on whether the memory controller is
reading or writing. In particular, the first input is selected when
the memory controller is reading data from the memory, and the
second input is selected when the memory controller is writing data
into the memory.
[0013] Other embodiments include a multiplexer having a first
input, a second input, and an output. The first input is configured
to receive the first data signal, and the second input is
configured to receive the second data signal. The output connects
to circuitry configured for placing either the first or second data
signal on the data line for writing into the memory.
[0014] Yet other embodiments, include a multiplexer having a first
input, a second input, and an output. The first input is configured
to receive the first data signal. The second input is configured to
receive a signal that has been read from memory and that has been
delayed. The output connects to circuitry that, when enabled, is
configured for transmitting, to the memory, either the first data
signal or the signal that has been read from memory.
[0015] Other embodiments include both a first and second
multiplexer. The first multiplexer includes an output, a first
input, and a second input. The output of the first multiplexer is
connected to an input of the data de-skewer. The first input of the
first multiplexer is connected to the data line to receive a signal
indicative of a bit that has been read from the memory. The second
input of the first multiplexer is configured to receive the first
data signal. The de-skewer has an output that is selectively
connected to the data line so that the second data signal can be
placed on the data line. The first input of the first multiplexer
is selected when the memory controller is reading data from the
memory, and the second input of the first multiplexer is selected
when the memory controller is writing data into memory. The second
multiplexer includes a first input, a second input, and an output.
The first input of the second multiplexer is configured to receive
the first data signal. The second input of the second multiplexer
is configured to receive the second data signal. The output of the
second multiplexer connects to circuitry configured for placing
either the first data signal or the second data signal on the data
line for writing into the memory.
[0016] In another aspect, the invention features a method for
controlling a write to a memory. Such a method includes receiving a
first data signal that has been read from the memory, at a first
de-skewer, applying a read-compensation skew to the first data
signal, receiving a second data signal that represents a bit of a
byte that is to be written to the memory, selecting a
write-compensation skew, and at the first de-skewer, skewing the
second data signal by the selected write-compensation skew, thereby
generating a third data signal. Selecting the write-compensation
skew includes selecting a skew that increases a likelihood that the
third data signal will be sampled during a data-valid window
thereof.
[0017] In some practices, selecting the write-compensation skew
includes selecting the write-compensation skew to maximize an
extent to which all third data signals associated with the byte
have overlapping data-valid windows.
[0018] In other practices, selecting the write-compensation skew
includes determining a compensation skew that, when added to an
inherent skew in the data signal, results in a constant that is the
same for all of the second data signals associated with the
byte.
[0019] Other practices include, for each of a set of candidate
compensation skews, applying the candidate compensation skew to
data, writing the data into memory, reading the data back from
memory, determining that the data read back from memory fails to
match the data that was written into memory, and based on the
determination, removing the compensation skew from the set of
candidate compensation skews.
[0020] In another aspect, the invention features an apparatus for
controlling memory. Such an apparatus includes a memory controller,
and an interface to data lines connecting the memory controller to
the memory. Each of the data lines carries a signal that
corresponds to a bit of a byte that is to be written to the memory.
The interface includes, for each of the data lines, a transmitter
to drive bits to be written into the memory onto a data line
coupled to the transmitter. The interface includes, for each of the
data lines, a receiver to detect bits that have been read from the
memory from a data line coupled to the receiver. The memory
controller includes, for each of the data lines, a data de-skewer
used in a writing mode and a reading mode. For each of the data
lines, the data de-skewer is configured to receive a first data
signal. Each of the data lines is associated with an inherent skew.
The data de-skewer applies a compensation skew to the first data
signal to generate a second data signal. In the writing mode the
first data signal is a signal that represents a bit that is to be
written into the memory, and in the reading mode the second data
signal is a signal that represents a bit that has been read from
the memory.
[0021] Among the embodiments are those in which, for each of the
data lines, the compensation skew applied to the first data signal
to generate a second data signal is obtained by training the data
de-skewer in the reading mode. These include embodiments in which,
for each of the data lines, the data de-skewer uses the
compensation skew, which was obtained by training the data
de-skewer in the reading mode, for skewing a data signal that
represents a bit that is to be written into memory. These also
include embodiments in which the data de-skewer is trained in the
reading mode using data signals that represent bits provided the
memory during normal operation thereof.
[0022] Also among the embodiments are those in which the data
de-skewer is trained based on locating timing of a
leading-transition interval and trailing-transition interval of
data signals representing bits propagating over each data line. In
some of these embodiments, locating timing of a leading-transition
interval and a trailing-transition interval of a first data signal
includes oversampling the first data signal based on a clock signal
corresponding to the first data signal. In others, oversampling the
first data signal based on the corresponding clock signal includes
sampling the first data signal based on the corresponding clock
signal and multiple time-shifted versions of the corresponding
clock signal.
[0023] In other embodiments, for each of the data lines, the data
de-skewer uses a first input value for determining the compensation
skew in the reading mode, and uses a second input value, different
from the first input value, for determining the compensation skew
in the writing mode. Among these embodiments are those in which,
for each of the data lines, the second input value is obtained by
training the data de-skewer by writing test data into the memory in
the writing mode and reading the test data from the memory in the
reading mode using the compensation skew determined in the reading
mode. These embodiments include those in which, for each of the
data lines, the first input value is obtained by training the data
de-skewer based on locating timing of a leading-transition interval
and trailing-transition interval of a data signal representing bits
propagating over the data line.
[0024] In another aspect, the invention features a method for
controlling a memory. Such a method includes causing a data
de-skewer to operate in a writing mode; at the data de-skewer,
receiving a first signal, and skewing the first data signal by a
first compensation skew; causing the data de-skewer to operate in a
reading mode; at the data de-skewer, receiving a second signal, and
skewing the second signal by a second compensation skew. In this
method, the first signal represents a bit from a byte that is to be
written to the memory, and the second signal represents a bit from
a byte that has been read from the memory. Thus, the same de-skewer
is being used in both directions.
[0025] Some practices include training the de-skewer while the
memory is operating in the reading mode. Among these are practices
in which training the de-skewer comprises generating the first
compensation skew, those in which training the data de-skewer in
reading mode comprises using data signals that represent bits
provided by the memory during normal operation of the memory, and
those in which training the data de-skewer in reading mode
comprises locating a leading-transition interval and a
trailing-transition interval of a data signal that represents a bit
that is propagating over a data line.
[0026] In those practices that involve locating the
leading-transition interval and the trailing-transition interval of
the data signal, these steps include oversampling the data signal
based on a clock signal, wherein the clock signal is a clock signal
that corresponds to the data signal. In some of these practices,
oversampling the data signal based on the corresponding clock
signal comprises sampling the first data signal based on a
corresponding clock signal and on multiple time-shifted versions of
the corresponding clock signal.
[0027] Alternative practices of the method include those that
include, for each of a plurality of data lines, causing the data
de-skewer to use a first input value for determining a compensation
skew in the reading mode, and causing the data de-skewer to use a
second input value, different from the first input value, for
determining a compensation skew in the writing mode. Among these
are practices in which, for each of the data lines, obtaining the
second input value by training the data de-skewer. In these
practices, training the data de-skewer comprises writing test data
into the memory in the writing mode, and reading the test data from
the memory in the reading mode using the compensation skew
determined in the reading mode. Some of these practices further
include, for each of the data lines, obtaining the first input
value by training the data de-skewer based on locating timing of a
leading-transition interval and trailing-transition interval of a
data signal representing bits propagating over the data line.
[0028] Some additional practices include selecting the first
compensation skew and the second compensation skew to be equal to
each other, and alternatively, selecting the first compensation
skew and the second compensation skew to be different from each
other.
[0029] The techniques described herein enable compensation of
systematic skew that may exist between a memory and a memory
controller. This may improve the setup and hold time margin for a
high speed memory interface with minimal additional hardware by
re-using the same compensation circuitry for de-skew of signals in
both directions over bidirectional data lines. The interface
includes transmitter circuitry for driving signals onto
electrically conducting pathways of the data lines propagating
toward the memory, and receiver circuitry for detecting signals
propagating in the opposite direction from the memory over those
same pathways. For example, the interface includes, for each data
line, a transmitter and receiver pair that can be alternatively
selected for transmission and reception, respectively. In addition
to re-use of the compensation circuitry that compensates the delay
skew difference among data lines, trained compensation settings can
be re-used to enable dynamic re-training for both reading and
writing, as described in more detail below.
[0030] These and other features of the invention will be apparent
from the following description and the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
[0031] FIG. 1 shows several bits overlaid on each other to show a
data-valid window.
[0032] FIG. 2 shows a flow chart for carrying out a write-skewing
procedure to ensure sampling during the data-valid window shown in
FIG. 1.
[0033] FIG. 3 shows sweeping across compensation skew settings to
identify optimal compensation skew settings to enlarge the
data-valid window shown in FIG. 1.
[0034] FIG. 4 shows the derivation of a compensation vector to be
used for compensation of the bits in FIG. 1.
[0035] FIG. 5 shows a circuit for providing a strobe according to
the procedure set forth in FIG. 2.
[0036] FIG. 6 shows a circuit for providing data signals according
to the procedure set forth in FIG. 2.
DETAILED DESCRIPTION
[0037] FIG. 1 shows a superposition of bits from a data unit after
having been placed on respective data lines by the controller and
having reached a memory. The particular example shows a data unit
with eight bits. However, the principles described herein are
independent of the actual number of bits.
[0038] The superposition has three intervals: a leading-transition
interval 10 during which the bits are potentially transitioning
between different values, a data-valid window 12 during which all
bits are held at their correct values, and a trailing-transition
interval 14 during which the bits are also potentially
transitioning between different values.
[0039] In the example of FIG. 1, for illustration purposes, all n
bits happen to transition from low to high and then from high back
to low, corresponding to a bit sequence of 010 for all n bits.
Also, the terms "leading" and "trailing" are relative to a time
slot for a particular bit. The trailing-transition interval of one
bit is the same as the leading-transition interval of the
subsequent bit.
[0040] It is apparent from the FIG. 1 that the individual bits have
been skewed. As a result, the extent of each of these intervals is
not the same across all data lines. Some data lines reach their
data-valid window 12 earlier than others. Other data lines may
start their trailing-transition interval 14 sooner than most. This
tends to broaden the aggregate leading-transition interval 10 and
the aggregate trailing-transition interval 14 of the collection of
bits as a whole, and to do so at the expense of a now shorter
aggregate data-valid window 12.
[0041] Clearly, the best time for a strobe to sample valid values
for all of the bits is while all n data lines are still in their
respective data-valid windows 12. Timing the strobe this way will
ensure that all bits are in a state suitable for sampling.
[0042] Given the technological limits on how short the transitions
of the leading-transition interval 10 and trailing-transition
interval 14 can be, as clock rates have become faster, the
data-valid window 12 has become shorter. Thus, it has become easier
for the strobe to miss the data-valid window 12. This can result in
inadvertently sampling one or more bits during the
leading-transition or trailing-transition interval 10, 14 when a
signal is not guaranteed to be at its correct value.
[0043] A solution to this difficulty is to de-skew the individual
data signals so as to reduce the extent of the aggregate
leading-transition and trailing-transition intervals 10, 14, thus
extending the aggregate data-valid window 12. This would be
followed by adding a skew that is required by the receiving memory
between all n data bits and the strobe. The extent of the relative
skew is selected to maximize the likelihood of having the memory
sample the data lines during the extended aggregate data-valid
window 12 thus created. In the preferred embodiment, the relative
skew is caused by skewing the data relative to a fixed strobe.
[0044] A signal's timing can be controlled by passing the signal
through a delay line. A delay-line has multiple delay stages
u.sub.1, u.sub.2, . . . u.sub.N, each of which yields an output
that is a delayed version of the input. The extent of this delay
monotonically increases with the index 1, 2, . . . N of the delay
stage. To delay the signal by a particular amount, one selects an
appropriate delay stage and obtains the delayed signal from the
final output.
[0045] Each one of the data signals shown in FIG. 1 has an inherent
skew that governs when the data-valid window 12 for that data
signal begins. This skew varies from one bit to the next. Although
the inherent skew cannot easily be eliminated, it can nevertheless
be compensated for.
[0046] The method and apparatus disclosed herein de-skews the
individual data signals shown in FIG. 1 by identifying a set of
compensation skews such that the sum of a data signal's inherent
skew and its compensation skew is substantially the same for each
data signal. As a result, the aggregate data-valid window 12 will
increase and the aggregate leading-transition and
trailing-transition intervals 10, 14 will decrease. The process of
identifying these compensation skews is carried out during a
training procedure. This might occur, for example, when a computer
is first turned on or rebooted. Once the compensation skews for
each bit have all been determined, they are stored (e.g., in
registers) for use during an operating procedure.
[0047] In some embodiments, the process begins by passing the
strobe through a delay line and skewing it by an amount that is
about midway across the range of the delay line. In other words, if
the delay line applies a maximum delay of t.sub.max and a minimum
delay of t.sub.min, then the strobe is skewed by about
0.5*(t.sub.max-t.sub.min).
[0048] Referring now to FIG. 2, an example of a write-de-skewing
process includes a training procedure 16 and an operating procedure
18. The training procedure 16 breaks down into a measurement phase
20 and an assignment phase 22.
[0049] The measurement phase 20 includes, for each bit, sweeping
across a range of compensation skews to identify an optimal
compensation skew for that bit. In particular, the measurement
phase 20 includes choosing a particular compensation skew (step
24), applying that compensation skew, and comparing a bit value
written to the memory with a bit value read back from the memory
(step 26) to see if it results in a write error. After repeating
this procedure for all bits (step 28), a new compensation skew is
obtained by incrementing or decrementing the old one (step 30). The
procedure then begins again all over again with the new
compensation skew. This procedure can be performed for multiple
bytes of a memory interface concurrently. For example, bit 0 of all
bytes can be tested at the same time, then bit 1 of all bytes can
be tested, etc. Also, in some embodiments, instead of repeating the
procedure sequentially for each bit of a byte, the error for all
bits of all bytes is analyzed in parallel.
[0050] In more detail, the measurement phase 20 includes, for each
value of compensation skew, causing each bit to be written into
memory and then reading the bit back. If the value read back is
consistent with the value written, then that compensation skew is
identified as one of several possible choices for the optimal
compensation skew that could be applied to that bit. This process
is carried out in parallel for all bits. Expressed in pseudo-code,
an example of this procedure is as follows:
TABLE-US-00001 FOR I=1..NBITS FOR J=1..NSKEWS TIME=0 AT
(TIME=J*SKEW_STEP) WRITE(BIT(I),MEMORY_LOCATION)
X=READ(MEMORY_LOCATION) IF (X==BIT(I)) THEN WRITE(J,GOOD_LIST) NEXT
J NEXT I
[0051] The outer "FOR" loop is only shown for clarity. In practice,
this procedure is carried out in parallel for all bits.
[0052] The assignment phase 22 includes examining the set of
candidate passed compensation skews to determine the optimal
compensation skew for each bit (step 32). These optimal
compensation skews are then assigned to each bit (step 34).
[0053] At this point, the training procedure 16 is complete.
[0054] FIG. 3 illustrates the process of sweeping across the
various delay settings for a first bit, as discussed in connection
with the measurement phase (step 20). The topmost waveform shows a
strobe 36 that has been delayed by the strobe skew defined above.
The remaining eight waveforms correspond to eight settings on a
delay line.
[0055] As is apparent, in a first skew setting 38, in which the
data signal has sustained a compensation skew of of the total
compensation range of the delay line, the rising edge of the
strobe, in this example, is close to the trailing-transition
interval 14. In the seventh, and eight skew settings 40, 42, where
the data signal has sustained compensation skews of 75% and 87.5%
of the total compensation range of the delay line, the rising edge
of the strobe is now close to the leading-transition interval 10.
Both of these skew settings result in reading a value that is
inconsistent with what was written. According to FIG. 3, the best
compensation skew to apply to the first bit would be the fourth
skew setting 44, which is 37.5% of the total compensation range of
the delay line.
[0056] The fact that the optimal skew setting for the first bit is
the fourth skew setting 44 says nothing about what the optimal skew
settings might be for the other seven bits. In fact, FIG. 4 shows
an example of the optimal skew settings for the first bit and the
remaining seven bits. As is apparent, many of the bits are
optimized by the fourth skew setting 44. However, there are some
outliers. For instance, bit 7 is optimized by the second skew
setting.
[0057] The result of the training procedure 16 is thus an
eight-dimensional optimal compensation vector (or "de-skewing
vector"), which in the case shown in FIG. 4 is [4 4 4 5 4 3 3 2].
The optimal compensation vector is then saved and applied for use
in operating procedure 18.
[0058] During the operating procedure 18, normal reading and
writing operations occur but with the write-skew feature enabled
(step 46). The compensation skew for each bit is stored in a
register associated with that bit. The compensation skew for a
particular bit is then applied to that bit every time a write is to
occur. Then, once the bits have all been placed on their respective
data lines, the strobe is placed on the strobe line after the lapse
of the strobe skew.
[0059] A method and apparatus for using a de-skewer for de-skewing
data signals that are being read from a memory is described in U.S.
patent application Ser. No. 14/046,879, filed on Oct. 4, 2013, the
contents of which are herein incorporated by reference. The method
and apparatus disclosed herein re-uses the same de-skewer that is
used for de-skewing data signals that are being read from the
memory for de-skewing data signals that are to be written to the
memory. The de-skewer can be trained separately for reading and
writing. Alternatively, a training re-use mode has the advantage of
enabling a de-skewing vector determined for reading to be re-used
for writing as well as reading, as described in more detail
below.
[0060] FIG. 5 shows a strobe-communication circuit 48 for
communicating strobes between a memory controller 50 and a memory
52 via a bidirectional strobe line 54. The particular circuit shown
has two modes of operation: a read mode, in which it receives a
read-strobe from the memory 52, and a write mode, in which it sends
a write-strobe to the memory 52.
[0061] The write mode can be a conventional write-mode or a skewed
write-mode. In the conventional write-mode, the write-strobe is
delayed by a nominal amount that is selected in an attempt to
sample at the midpoint of a data-valid window 12. A write-strobe
having this property will be referred to as a "nominal
write-strobe." In the skewed write-mode, the write-strobe is first
passed through a de-skewer that delays it by some predetermined
amount that depends on the minimum and maximum delay of the
de-skewer. This kind of write strobe will be referred to as a
"skewed write-strobe."
[0062] The strobe-communication circuit 48 includes a transmitter
56 for placing a write-strobe on the strobe line 54. This
write-strobe is to be provided to the memory 52 during a write
operation.
[0063] A first multiplexer 58 has an output connected to the
transmitter 56. The output of the first multiplexer 58 provides the
write-strobe that is to be placed on the strobe line 54 by the
transmitter 56. The transmitter 56 is, in some embodiments,
configured to convert logic levels used by the memory controller 50
into logic levels used by the memory 52.
[0064] The first multiplexer 58 has first and second inputs 60, 62
for receiving two different versions of a write-strobe. The first
input 60 receives a nominal write-strobe. The second input 62
receives a skewed write-strobe. The choice of whether to place the
skewed write-strobe or the nominal write-strobe on the strobe line
54 is made by control logic 64 connected to a control input 66 of
the first multiplexer 58.
[0065] The strobe-communication circuit 48 further includes a
receiver 68 for receiving a read-strobe on the strobe line 54
during a read operation. The receiver 68 connects to a first input
70 of a second multiplexer 72, a second input 74 of which receives
the nominal write-strobe. The receiver 68 is, in some embodiments,
configured to convert logic levels used by the memory 52 into logic
levels used by the memory controller 50.
[0066] The second multiplexer 72 has an output 76 that leads to a
de-skewer 78. Thus, the de-skewer 78 can receive either the nominal
write-strobe or a read-strobe. The choice between these two is made
by a signal provided by the control logic 64 to a control input 80
of the second multiplexer 72.
[0067] When being used in connection with writing into memory, a
de-skewer output 82, or the second input of the first multiplexer
58, provides a skewed strobe to the second input 62 of the first
multiplexer 58. On the other hand, when being used in connection
with reading from memory, the de-skewer output 82 provides the
skewed strobe to whatever entity would normally use the
read-strobe. The amount of skew may be fixed, or may be tuned based
on a range of skew values determined for the data signals, as
described below with reference to FIG. 6. In some embodiments, the
amount of skew is fixed halfway between the minimum and maximum
possible skews provided by the de-skewer 78.
[0068] When carrying out a read operation, the receiver 68 is
active and the transmitter 56 is inactive. The control logic 64
selects the first input 70 of the second multiplexer 72. As a
result, the read-strobe passes through the de-skewer 78, which then
delays it by some predetermined amount. The skewed read-strobe also
happens to be provided to the second input of the first multiplexer
58. However, since the transmitter 56 is inactive, this has no
effect.
[0069] When carrying out a normal write-operation, the receiver 68
is inactive and the transmitter 56 is active. The control logic 64
selects the first input 60 of the first multiplexer 58. As a
result, the nominal write-strobe is placed on the strobe line
54.
[0070] When carrying out a skewed write-operation, the receiver 68
is inactive and the transmitter 56 is active. The control logic 64
selects the second input 62 of the first multiplexer 58 and the
second input 74 of the second multiplexer 72. This means that the
signal placed on the strobe line 54 will be the output of the
de-skewer 78. As it turns out, since the second input 74 receives
the nominal write-strobe, the output of the de-skewer 78 will be a
skewed write-strobe. As a result, a skewed write-strobe is placed
on the strobe line 54.
[0071] The strobe communication circuit 48 thus offers the
advantage of re-using the de-skewer 78 for both read and write
operations. This saves considerable die space and reduces overall
power consumption.
[0072] FIG. 6 shows one of eight data-communication circuits 84 for
communicating data between a memory controller 50 and a memory 52
via a bidirectional data line 86. The particular circuit shown has
two modes of operation: a read mode, in which it receives a data
bit from the memory 52, and a write mode, in which it sends a data
bit to the memory 52.
[0073] The write mode can be a conventional write-mode or a skewed
write-mode. In the conventional write-mode, the data bit to be
written is not delayed on purpose. A data bit having this property
will be referred to as a "nominal data-bit." In the skewed
write-mode, a de-skewer has delayed the data bit to be written in
an attempt to maximize the data-valid window that can be sampled by
the strobe. This kind of data bit will be referred to as a "skewed
data bit."
[0074] The data-communication circuit 84 includes a transmitter 88
for placing a data bit on the data line 86. This data bit is to be
provided to the memory 52 during a write operation. The transmitter
88 is, in some embodiments, configured to convert logic levels used
by the memory controller 50 into logic levels used by the memory
52.
[0075] A first multiplexer 90 has an output connected to the
transmitter 88. The output of the first multiplexer 90 provides the
data bit that is to be placed on the data line 86 by the
transmitter 88.
[0076] The first multiplexer 90 has first and second inputs 92, 94
for receiving two different versions of a data bit. The first input
92 receives a nominal data-bit. The second input 94 receives a
skewed data-bit. The choice of whether to place the skewed data-bit
or the nominal data-bit on the data line 86 is made by control
logic 96 connected to a control input 98 of the first multiplexer
90.
[0077] The data-communication circuit 84 further includes a
receiver 100 for receiving a data bit on the data line 86 during a
read operation. The receiver 100 connects to a first input 102 of a
second multiplexer 104, a second input 106 of which receives the
nominal data-bit. The receiver 100 is, in some embodiments,
configured to convert logic levels used by the memory 52 into logic
levels used by the memory controller 50.
[0078] The second multiplexer 104 has an output 108 that leads to a
de-skewer 110. Thus, the de-skewer 110 can receive either the
nominal data-bit that is to be written into memory or a data bit
that has been read from memory. The choice between these two is
made by a signal provided by the control logic 96 to a control
input 112 of the second multiplexer 104.
[0079] When used in connection with writing data into memory, a
de-skewer output 114 provides a skewed data signal to the second
input 94 of the first multiplexer 90. When used in connection with
reading data from memory, the de-skewer output 114 provides the
skewed data signal to whatever entity requested the data to be
read.
[0080] As discussed in connection with FIG. 2, the amount of skew
is determined during the training procedure 16, which may be
initially performed before the operating procedure 18. Some
practices have separate read-training and write-training
procedures. During the read-training procedure, optimal
compensation skews for use in connection with reading from memory
52 are determined. During the write-training procedure, optimal
compensation skews to be used in connection with writing into
memory 52 are determined.
[0081] Preferably, the read-training procedure is performed first,
followed by the write-training procedure. This means that the
optimal compensation skews derived from the read-training procedure
can be used for reading values during the comparison step 26. Thus,
results of the write-training procedure depend at least in part on
results of the read-training procedure.
[0082] Some implementations have other differences between
read-training and write-training. For example, the write-training
procedure 16 described above is performed based on a round-trip
comparison between a value written to memory 52 and a value read
from memory 52. In contrast, the read-training procedure can be
performed by simply making use of any data pattern available from
the memory 52. This is because the data-communication circuit 84
does not need to write a value and then read it back in for
comparison. Instead, read-training can be carried out by reading
the data pattern and locating the timing of the actual
leading-transition interval 10 and trailing-transition interval 14
for each bit over the compensation range of the delay line. This
can be done by oversampling the data signal to locate leading and
trailing edges of a single bit, and use the specific locations of
those edges to determine the optimal compensation skew values. The
oversampling can be performed based on the incoming strobe signal
corresponding to the data signal. For example, the oversampling may
include sampling using multiple time-shifted versions of the strobe
signal, as described in more detail in U.S. patent application Ser.
No. 14/046,879.
[0083] In some implementations, the read-training is initially
performed before normal operation begins, using a pre-defined test
data pattern written into the memory, or provided by the memory as
a built-in pattern. Additionally, the read-training procedure can
also be carried out during normal operation of the memory using any
data pattern that happens to be read from the data lines. An
advantage of being able to carry out read-training during normal
operation by making use of any data pattern is the ability to
periodically re-train during operation without needing to suspend
normal operations between the memory controller 50 and the memory
52. This re-training may be useful to obtain updated compensation
skew values that have changed due to a temperature and voltage
change, for example.
[0084] When carrying out a read operation, the receiver 100 is
active and the transmitter 88 is inactive. The control logic 96
selects the first input 102 of the second multiplexer 104. As a
result, the data bit that has just been read passes through the
de-skewer 110, which then delays it by some predetermined amount.
The skewed data bit that has just been read also happens to be
provided to the second input of the first multiplexer 90. However,
since the transmitter 88 is inactive, this has no effect.
[0085] When carrying out a normal write-operation, the receiver 100
is inactive and the transmitter 88 is active. The control logic 96
selects the first input 92 of the first multiplexer 90. As a
result, the nominal data-bit is placed on the data line 86.
[0086] When carrying out a skewed write-operation, the receiver 100
is inactive and the transmitter 88 is active. The control logic 96
selects the second input 94 of the first multiplexer 90 and the
second input 106 of the second multiplexer 104. This means that the
signal placed on the data line 86 will be the output of the
de-skewer 110. As it turns out, since the second input 92 receives
the nominal data-bit, the output of the de-skewer 110 will be a
skewed data-bit. As a result, a skewed data-bit is placed on the
data line 86.
[0087] The data-communication circuit 84 thus offers the advantage
of re-using the de-skewer 110 for both read and write operations.
This saves considerable die space and reduces overall power
consumption.
[0088] The control logic 96 also has the ability to control what
compensation skews are used for read operations and write
operations. In a compensation re-use mode, the control logic 96
selects the compensation skews determined during the read-training
procedure for both read operations and write operations. Because
much of the route used by the data lines as they traverse the
hardware between the memory controller 50 and the memory 52 is
bidirectional, path length difference for reading and writing will
be substantially identical. This means that the compensation skews
for reading can also provide improved setup and hold time margin
when they are re-used for writing. There may be some difference
between the optimal compensation skews for reading and writing,
which come from differences in the circuitry of the data lines that
is not in common for both transmission and reception, but the
reading compensation skews may be adequate for writing in some
cases.
* * * * *